Releases: 1HelloMan1/dsh-vision-fallback
Releases · 1HelloMan1/dsh-vision-fallback
Release list
v0.9.1 — Marketplace readiness
- Added the dsh-plugin npm keyword and a regression test for the installable DSH bundle manifest.
- Expanded documentation for compaction reuse, tool-result images, persistent observations, and restart recovery.
- Published as dsh-vision-fallback@0.9.1 on npm.
v0.9.0 — 压缩复用与会话位置观察缓存
- 新增压缩(compaction)调用线路:优先复用主模型已有视觉观察,不再整批重复识图
- 观察缓存持久化到 observations.json,服务重启和事件回放后可继续复用
- 缓存按会话、消息位置、图片和视觉路由隔离;同图在新消息位置会重新识别
- 工具结果图片进入模型投影,同时保留界面原始图片
- usage.jsonl 记录耗时、成功/失败状态、图片信息与 token 用量
- 21 项单元测试覆盖普通对话、工具结果、压缩、缓存和用量记录
v0.7.0 — 视觉模型主模型不再被接管
新增 mode 配置(默认 auto):\n\n- auto:主模型自带视觉(kimi-k3 / mimo-v2.5 等)时图片原样交给它,不再转文字;纯文本模型仍自动接管\n- always:旧行为(总是转文字)\n- never:完全停用视觉桥\n\n安装:dsh plugin --profile web add dsh-vision-fallback@0.7.0
v0.6.0 — 视觉桥修复
- 替换表清理:消息离开当前视图后丢弃陈旧替换记录\n- 视觉上下文更完整:preprocess 使用完整模型视图消息\n- 单元测试 11 个全过\n\n安装:dsh plugin --profile web add dsh-vision-fallback
v0.5.0 — Silent Vision Enhancement
Initial public release.
- Silent vision bridge: chat-box images are auto-converted to text observations for text-only main models (DeepSeek V4 Flash/Pro, etc.)
- UI keeps the original image (model-only surface replacement via agent/pre-step)
- Configurable vision model / endpoint / credential ref / prompt through the Web settings page
- Loopback-only config route, LRU cache, abort & timeout handling
- 10 unit tests passing; English + Chinese README
Install: dsh plugin --profile web add dsh-vision-fallback (or from source checkout)